fix: reject NaN/+inf in numeric settings validation

[bodono]

Summary

• IEEE NaN comparisons always return false, so x <= 0 / x < 0 / x >= 2 let NaN slip through every numeric setting in SCS_init. The solver then runs to completion producing NaN iterates with no error surfaced. SCS core's own validate() has the same gap.
• scale=+inf additionally triggers a misleading "ScsWork allocation error!" from the linear-system factorization; rho_x=+inf silently produces NaN iterates.
• Add explicit isnan/isfinite guards in the validation block. scale, rho_x, alpha must now be finite; eps_* and time_limit_secs reject NaN but still accept +inf (semantically meaningful — disables that stopping criterion / time limit).

Repro (before)

sol = scs.solve(data, {"q": [], "l": 2}, scale=float("nan"))
# -> runs to "max_iters reached", x=[nan]
sol = scs.solve(data, {"q": [], "l": 2}, scale=float("inf"))
# -> ValueError: ScsWork allocation error!  (misleading)

After

ValueError: scale must be a positive finite number

Test plan

• 11 new parametrized tests covering NaN on all 7 float settings, +inf rejection on scale/rho_x/alpha, +inf acceptance on eps_*/time_limit_secs
• Existing test_scale_zero_raises / test_rho_x_zero_raises regex relaxed to accommodate the updated error string
• Full suite green (371 passed, 66 skipped)

🤖 Generated with Claude Code